Fix i386 fdim double rounding (bug 20255).

fdim suffers from double rounding on i386 because subtracting two
double values can produce an inexact long double value exactly half
way between two double values.  This patch fixes this by creating an
i386-specific version of fdim - C, based on the generic version,
unlike the previous .S version - which sets the x87 precision control
to double precision for the subtraction and then restores it
afterwards.  As noted in the comment added, there are no issues of
double rounding for subnormals (a case that setting precision control
does not address) because subtraction cannot produce an inexact result
in the subnormal range.

Tested for x86_64 and x86.

	[BZ #20255]
	* sysdeps/i386/fpu/s_fdim.c: New file.  Based on math/s_fdim.c.
	* math/libm-test.inc (fdim_test_data): Add another test.

3 files changed, 57 insertions, 0 deletions

diff --git a/ChangeLog b/ChangeLog
index b6d20dc89d..a4a1cea4ab 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,9 @@
 2016-06-14  Joseph Myers  <joseph@codesourcery.com>
 
+	[BZ #20255]
+	* sysdeps/i386/fpu/s_fdim.c: New file.  Based on math/s_fdim.c.
+	* math/libm-test.inc (fdim_test_data): Add another test.
+
 	[BZ #6796]
 	[BZ #20255]
 	[BZ #20256]
diff --git a/math/libm-test.inc b/math/libm-test.inc
index cc41c63e7a..d6cad6634c 100644
--- a/math/libm-test.inc
+++ b/math/libm-test.inc
@@ -7203,6 +7203,9 @@ static const struct test_ff_f_data fdim_test_data[] =
     TEST_ff_f (fdim, 0, 9, 0, NO_INEXACT_EXCEPTION|ERRNO_UNCHANGED),
     TEST_ff_f (fdim, -9, 0, 0, NO_INEXACT_EXCEPTION|ERRNO_UNCHANGED),
     TEST_ff_f (fdim, 0, -9, 9, NO_INEXACT_EXCEPTION|ERRNO_UNCHANGED),
+#if TEST_COND_binary64
+    TEST_ff_f (fdim, 1.0, -0x1.0000000000001p-53, 1.0, INEXACT_EXCEPTION|ERRNO_UNCHANGED, 0x1.0000000000001p0, INEXACT_EXCEPTION|ERRNO_UNCHANGED, 1.0, INEXACT_EXCEPTION|ERRNO_UNCHANGED, 0x1.0000000000001p0, INEXACT_EXCEPTION|ERRNO_UNCHANGED),
+#endif
     TEST_ff_f (fdim, min_subnorm_value, min_subnorm_value, 0, NO_INEXACT_EXCEPTION|ERRNO_UNCHANGED),
     TEST_ff_f (fdim, -min_subnorm_value, -min_subnorm_value, 0, NO_INEXACT_EXCEPTION|ERRNO_UNCHANGED),
     TEST_ff_f (fdim, min_subnorm_value, -min_subnorm_value, 2*min_subnorm_value, NO_INEXACT_EXCEPTION|ERRNO_UNCHANGED),
diff --git a/sysdeps/i386/fpu/s_fdim.c b/sysdeps/i386/fpu/s_fdim.c
new file mode 100644
index 0000000000..e4db9bf66c
--- /dev/null
+++ b/sysdeps/i386/fpu/s_fdim.c
@@ -0,0 +1,50 @@
+/* Return positive difference between arguments.  i386 version.
+   Copyright (C) 1997-2016 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <errno.h>
+#include <fpu_control.h>
+#include <math.h>
+#include <math_private.h>
+
+double
+__fdim (double x, double y)
+{
+  if (islessequal (x, y))
+    return 0.0;
+
+  /* To avoid double rounding, set double precision for the
+     subtraction.  math_narrow_eval is still needed to eliminate
+     excess range in the case of overflow.  If the result of the
+     subtraction is in the subnormal range for double, it is exact, so
+     no issues of double rounding for subnormals arise.  */
+  fpu_control_t cw, cw_double;
+  _FPU_GETCW (cw);
+  cw_double = (cw & ~_FPU_EXTENDED) | _FPU_DOUBLE;
+  _FPU_SETCW (cw_double);
+  double r = math_narrow_eval (x - y);
+  _FPU_SETCW (cw);
+  if (isinf (r) && !isinf (x) && !isinf (y))
+    __set_errno (ERANGE);
+
+  return r;
+}
+weak_alias (__fdim, fdim)
+#ifdef NO_LONG_DOUBLE
+strong_alias (__fdim, __fdiml)
+weak_alias (__fdim, fdiml)
+#endif